Agent Capability Is Outrunning Agent Judgment

In one week, agents showed both their ceiling and their floor

Four AI agents running on Google's Gemini 4 Argon beat Minecraft without any human help, slaying the Ender Dragon in a fully autonomous run. With no predefined roles or strategies, they coordinated on their own, wrote 2,783 programs, exchanged 229 chat messages, and died 605 times.

In the same batch of signals, a second item matters more for product teams: a new benchmark called Taste-Bench tests how well agents exercise strategic taste during long-running tasks — picking the right direction at key decision points. Even the best model, GPT-5.6 Sol, picks the right path on just under 60 percent of simple binary questions. The sharper finding: thinking longer doesn't help.

Read together, the two signals describe the current reality of agent deployment: execution chains can be long and stable, while the sense of direction stays poor.

This is a taste problem, not an alignment problem

Taste-Bench does not measure safety or factual accuracy. It measures trade-offs: among several viable paths, which one is worth taking over the long run. Just under 60 percent means nearly four in ten critical forks get the wrong direction; and "thinking longer doesn't help" suggests this is not a reasoning-budget issue but a missing evaluation standard and preference model.

For agent builders the implication is direct: extending chains of thought and adding reflection rounds to lift long-task success rates has a ceiling. What can be fixed sits outside the model — turning judgment criteria into explicit rules, checkpoints, and human review points rather than expecting the model to have taste on its own.

Norms and liability are being filled in from outside

Anthropic's move is a footnote to the same story: it is consulting with law firms to develop negotiation norms for AI agents that would haggle over prices, resolve complaints, or improve contracts on behalf of users, and it is hiring lawyers and humanities scholars as "Conceptual Reasoning Fellows" to improve Claude in these judgment areas.

The legal signal is harder. After a landmark Munich ruling on AI search liability, Google appealed and then quietly settled. The Aachen Regional Court has now applied the exact same reasoning in a new case, where Google's AI falsely attributed criminal offenses to a doctor who was not actually involved. Both courts land in the same place: Google is liable for what its AI overviews say.

In other words, an agent's taste deficit is not only a UX problem. It becomes a legal liability problem.

The organizational layer: junior roles are shrinking

Anthropic's own workforce data offers a third angle: the share of junior positions among new hires dropped from 27 to 9 percent since early 2025, and by August 2026 AI agents across the company were already logging six times as many hours as humans. Researchers warn of the consequences. TechCrunch adds that Anthropic's workforce has become less junior while employees spend more time in meetings.

This closes the loop with the technical signals: agents absorb large volumes of replicable execution work, so entry-level roles get compressed; meanwhile judgment, norm-setting, and cross-party coordination — the non-execution work — grow heavier, showing up as more meetings and a more senior mix. The risk is that junior roles are precisely where judgment gets trained.

Edge signals: data boundaries and asset reshuffling

Meta removed language from its terms that had raised scraping and AI training concerns; a spokesperson insisted the update did not reflect a change in the way Meta crawls the web. Wording changes are not behavioral commitments, so this stays worth watching.

Disney agreed to hire the majority of Character.AI's workforce, and a filing says the "vast majority" of employees who were offered jobs accepted. However, this is a stalking horse bid, so another company could still come forward with a higher offer.

Conclusion: treat judgment as an engineering problem

Four actionable steps for readers:

1. Upgrade evaluation. Beyond task completion rates, add long-horizon, multi-fork decision tests (Taste-Bench is a useful reference) and log which types of forks the errors cluster around.
2. Front-load the norms. For any agent that talks to the outside world, negotiates, or handles complaints, write boundary rules and human review points first — do not wait for an incident.
3. Design in liability. The two German courts show that what AI-generated content says externally is traced back to the operator; output review and auditable logs should be defaults.
4. Rebuild the path for developing judgment. If junior roles are compressed, replace learning-by-doing with reviewable real tasks and explicit retrospective mechanisms — otherwise in three to five years nobody will be able to take over norm-setting.

Sources

• AI agents have bad taste in long-term decisions and a new benchmark proves it — The Decoder
• Anthropic consults law firms to set fair negotiation norms before AI agents start cutting deals — The Decoder
• Two German courts now hold Google liable for what its AI overviews say — The Decoder
• Four Gemini 4 Argon agents beat Minecraft without any human help — The Decoder
• Anthropic has nearly stopped hiring junior staff because AI agents now work six times as many hours — The Decoder
• Anthropic says its workforce has become less junior and is spending more time in meetings — TechCrunch
• Meta removes language that raised scraping and AI training concerns — TechCrunch
• Disney agrees to hire the majority of Character.AI's workforce — TechCrunch